Title:Convergence Guarantees for Non-Convex Optimisation with Cauchy-Based Penalties

Abstract:In this paper, we propose a convex proximal splitting methodology with a non-convex penalty function based on the heavy-tailed Cauchy distribution. We first suggest a closed-form expression for calculating the proximal operator of the Cauchy prior, which then makes it applicable in generic proximal splitting algorithms. We further derive the required condition for minimisation problems with the Cauchy based penalty function that guarantees the convergence to the global minimum even though it is non-convex. Setting the system parameters by satisfying the proposed condition keeps the overall cost function convex, which is minimised via the forward-backward (FB) algorithm. The Cauchy regularisation based proposed method is evaluated in solving two generic signal processing examples, which are a 1D signal denoising in the frequency domain and two 2D image reconstruction studies of \textit{de-blurring} and \textit{denoising}. We experimentally verify the proposed convexity conditions for various cases, and show the effectiveness of the proposed Cauchy based non-convex penalty function over the state-of-the-art penalty functions of $L_1$ and total variation ($TV$) norms.